Local Edition

A weekly local edition for every store in a chain, built from the taste of each block (Qloo + NVIDIA Nemotron on Nebius).

Live demo: https://localedition.pages.dev

Try the sample chain in 30 seconds

1. Open https://localedition.pages.dev and press Open the sample chain. You land on a map of Kettle & Kiln, a sample coffee chain with eight stores at real New York corners.
2. Press Build this week's editions and watch the agent work store by store (tool calls, models, tokens and cost stream in live).
3. Open any store: the front page has four items with their evidence (open the evidence drawer). Press Approve all, then open the staff card.
4. Optional: Connections shows Square, Slack, Gmail/Outlook and the calendar feed; /proof and /limits show what ran, on which model, at what cost, and every budget.

What it does

For each store, once a week, the agent builds an edition from that block's taste:

• Special: one new drink or bite, taken from the chain's own menu (read from a Square catalog, a photo of the menu board, or a CSV), tuned to the block.
• Sound brief: what to play in the room, with artists that skew local.
• Local partner: a neighbouring place worth a tie-in, with an offer idea.
• Event night: something to host that week, tied to a real local event when one is found.

Every item carries evidence (source, endpoint, numbers) and an item without evidence is dropped, never shown. A Nemotron 3 Ultra reviewer checks the whole chain for near-identical stores. Head office approves on a map of all stores; approved editions become a one-page staff card (phone first, public link) for each store.

Connectors, and what is live

Each connector is built and tested against the providers' documented API shapes (196 unit tests across the Square, Slack, Gmail/Outlook and calendar code, with provider calls faked). The Composio auth configs for Gmail and Outlook exist. No live sandbox or test account is connected yet: there are no Square or Slack credentials on the deployed app and no Gmail consent, so on https://localedition.pages.dev none of the three has run against its real provider.

Tool	What it does here	Tested against	Live on localedition.pages.dev today
Square	Two ways: imports the menu (locations become stores); approving a special publishes it back; signed webhooks update the app; sales per special are read back. OAuth or a token.	Square's documented API shapes; sandbox connection pending.	No
Slack	Head-office digest with Approve and Regenerate buttons (signed interactions), staff cards to a store channel.	Slack's documented payloads and signing; test-workspace connection pending.	No
Gmail / Outlook (via Composio)	The staff card is emailed to the store; replies from managers are classified (sold out, sold well, change request, problem, other) and become suggestions for next week's build. Needs a consented mailbox.	Composio's documented API shapes; auth configs exist, no mailbox consented yet.	No
Calendar feed	A private .ics link per store with its event night. Needs no external account.	Unit tests on the feed.	Yes

Built on

Technology	What it is	What it does here
Qloo Taste AI	A cross-domain taste and affinity API.	The block profile: what the people near each store lean toward, and how that differs from the city. Endpoints used (host hackathon.api.qloo.com): GET /search (find brands the chain resembles), GET /v2/insights (block tags, nearby artists, brands and places, heatmap, rising taste), GET /v2/trending (what is rising), GET /v2/analysis/compare (two stores side by side). No customer or personal data is ever sent: only coordinates, entity ids and tag urns.
NVIDIA Nemotron 3.5 Super VL (nvidia/Nemotron-3.5-Super-VL)	A vision-language model with tool calling.	The agent loop that plans each edition and calls tools, and the menu-board photo reader.
NVIDIA Nemotron 3 Ultra (nvidia/Nemotron-3-Ultra-550b-a55b)	A large reasoning model.	The reviewer: judges every store's edition against the others and asks for a regenerate with a hint.
NVIDIA Nemotron 3.5 Lightning (nvidia/Nemotron-3_5-Lightning)	A fast, cheap model.	Copy for items and staff cards, Tavily event triage, the live self-check, and classifying email replies.
Qwen3 Embedding 8B (Qwen/Qwen3-Embedding-8B)	A text embedding model.	Matches menu items and nearby places to each block's taste, and finds near-identical items across stores (cosine clash).
Nebius Token Factory	OpenAI-compatible inference for open models.	Serves every model above.
NeMo Guardrails 0.24.1	NVIDIA's guardrails library.	Certifies every item (a custom health-claim rail, prompt-injection stripping, and NeMo's self check output rail on Lightning) and runs a red-team eval. It runs in a Nebius Serverless Job (jobs/, guardrails/) that is started from the CLI on demand, with npx tsx scripts/nebius-job.ts certify eval. It is not on a schedule: Cloudflare Workers cannot reach Nebius's control plane (it answers Cloudflare egress with a 522, see docs/NEBIUS-REST.md). Results are posted back signed and show on /proof as manual runs. The live path in the Worker (every head-office edit and every build) uses a TypeScript mirror of the regex rails plus the Lightning self-check, and marks items checkedBy: live.
Tavily	A search API for agents.	Search and extract for this week's local events near each store; crawl of the chain's own site to draft the brand voice.
Cloudflare	Workers, Pages, D1, KV, R2, Durable Objects, Workflows, cron.	Worker via OpenNext fronted by Pages; D1 for data; KV for provider caches; R2 for menu photos; Durable Objects for the spend meter, build lock and call ledgers; Workflows for builds (one retried step per store); cron every Monday 00:00 UTC for the weekly rebuild, which runs on Token Factory and Tavily only.

Architecture

flowchart LR
  V[Head office browser] --> P[Cloudflare Pages front]
  P -->|service binding| W[Worker: Next.js via OpenNext]
  W --> D1[(D1)]
  W --> KV[(KV cache)]
  W --> R2[(R2 menu photos)]
  W --> DO[Durable Objects: spend, locks, Qloo and Tavily ledgers]
  W --> WF[Workflows: EditionBuild, WeeklyRebuild]
  CRON[Cron: Monday 00:00 UTC] --> WF
  WF --> A[Agent loop]
  A -->|Super VL, Ultra, Lightning, Qwen3 Embedding| TF[Nebius Token Factory]
  A --> Q[Qloo Taste AI]
  A --> T[Tavily]
  A --> G[Live guardrails: regex rails + Lightning self-check]
  W <--> SQ[Square]
  W <--> SL[Slack]
  W <--> CO[Composio: Gmail / Outlook]
  CLI[Developer CLI, on demand] -->|nebius ai job| NJ[Nebius Serverless Job: NeMo Guardrails 0.24.1]
  NJ -->|signed results| W

Loading

Limits and budgets

• Spend cap: SPEND_CAP_USD (default $5) on Token Factory, enforced in one Durable Object with reserve-then-settle. Each visitor workspace gets a $0.40 sub-budget and each network a $0.80 daily cap; $1.50 is reserved for the demo workspace and the weekly job.
• Qloo call ledger: at most 300 live calls in total and 40 a day across all visitors, reserved before each call and never refunded, plus a 14-day response cache. Regenerate never reaches Qloo.
• Tavily ledger: one global ledger of 80 credits in total and 10 a day, results cached 7 days; live search is limited to the demo workspace's unmodified sample stores and chain setup.
• Rate limits per workspace and per IP on builds, regenerate, staff cards, edits, menu reads, brand search and brand-voice drafting (the table is on /limits; the code is src/server/limits-table.ts).
• At most 60 new workspaces an hour; workspace data expires.

Honest limitations

• Kettle & Kiln is sample data, labelled so in the app. It is not a customer.
• Square, Slack and Gmail/Outlook are not connected on the live app yet (see the connectors table); only the calendar feed runs without an account.
• Qloo: live-only in production. Every Qloo result comes from a live Qloo call or a saved real answer (KV, 14 days); replay, fixture and recording fallbacks are unreachable there, and a failed call shows a retry state, never sample data. Calls go through a hard ledger (2,000 through 15 Dec, 200 a day). Real Qloo responses are never committed. Fixtures under fixtures/qloo/ exist for unit tests and next dev only.
• NeMo Guardrails runs only when someone starts the job from the CLI; the weekly cron does not call Nebius.
• The first live Tavily prefetch of a week can take longer than the agent's 25 s tool timeout; the work finishes and is cached.
• docs/ holds working notes and smoke-test records from the build; docs/CONTRACT.md, docs/CLOUDFLARE.md and docs/DEPLOY.md are the reference ones.

Run locally

npm install
NEXT_PUBLIC_DATA_MODE=mock npm run dev   # UI on typed mock data, no server calls
npm run dev                              # DATA_MODE=app: real routes; providers default to replay/fixture without keys (never in production)

Open http://localhost:3000. Add NEBIUS_API_KEY, TAVILY_API_KEY and QLOO_API_KEY to .env.local and set TF_MODE, TAVILY_MODE, QLOO_MODE to live to use real providers; .env* is gitignored. scripts/deploy.sh and scripts/deploy-secrets.sh deploy to your own Cloudflare account and read credentials through a local ~/.secret-store/inject.sh helper, so adapt them to your own secret handling.

Tests

npm run typecheck && npm run lint
npx vitest run                  # 115 files, 1360 tests
npm run e2e                     # Playwright (starts next dev)
cd guardrails && python3 -m pytest -q   # NeMo rails: 115 passed, 1 skipped
python3 -m pytest -q tests/e            # job scripts: 12 passed

License

MIT, see LICENSE. Copyright 2026 JoelOnBeat.